Connectomics: mapping neurons from images
Read these first if they are new to you
- Segmentation: a label for every voxel Intermediate
- Shapes as point clouds Intermediate
The idea in one paragraph
Connectomics maps every neuron in a piece of brain and every connection between them. Scientists slice the tissue thinly and photograph it with an electron microscope. The photos form one huge 3D image. Programs then label each neuron's voxels, people proofread the labels, and programs find the connections. The result is a wiring diagram of the brain, called a connectome. This page is an overview: each step has its own unit.
Neurons are thin, branching cells joined at synapses
A neuron is a nerve cell, a cell that carries signals. It has a cell body, which holds its nucleus, and long branches. Branches called dendrites mostly receive signals. A branch called the axon sends them. A synapse is a contact point where one neuron passes a signal to another. A connectome is the full list of which neurons connect to which, through how many synapses.
Many branches are thinner than 200 nm, about the smallest detail an ordinary light microscope can show. That is why connectomics uses an electron microscope, which can show details of a few nanometers.
A connectome is built in six steps
| Step | What happens | Unit |
|---|---|---|
| 1. Slice and image | Preserved tissue is cut or worn away in layers from a few to about 50 nanometers thick, and each layer is photographed. | — |
| 2. Stack | The photos are lined up and stacked. The voxels are often deeper than wide, for example 8 × 8 × 40 nm. | 3D images made of voxels |
| 3. Segment | Programs give each voxel a label, first as small pieces and then as joined objects. | Segmentation, supervoxels |
| 4. Proofread | People, helped by software, fix the objects that were wrongly joined or cut. | Two kinds of error, proofreading |
| 5. Find synapses | A program marks each synapse and the two neurons it joins. | — |
| 6. Wiring diagram | Neurons become dots and synapses become lines. | Graphs |
The scale is what makes connectomics hard
The whole brain of an adult fruit fly has about 140,000 neurons. The count covers the central brain and the two optic lobes, which handle vision. It leaves out the nerve cord in the fly's body, which works like a spinal cord. A whole adult mouse brain has about 70 million neurons, and a whole adult human brain about 86 billion.
Even a small piece of brain makes a huge image. One cubic millimeter of tissue, imaged at 8 × 8 × 40 nm per voxel, has 125,000 × 125,000 × 25,000 ≈ 3.9 × 1014 voxels. At one byte per voxel, that is about 390 terabytes.
No one can label that by hand, so programs do the segmenting. Programs make errors, and at this scale even a small error rate means many errors. So most of the human effort in connectomics goes into proofreading, and much of the research goes into software that finds errors for people to check.
Programs reason about neuron shapes as point clouds
Once each neuron piece has a label, a program can treat the piece as a shape. A common way is to sample points on its surface, giving a point cloud. Another is a skeleton: a set of connected center lines that follows each branch. Shapes let a program ask questions the raw voxels make hard, such as "does this cut-off branch continue in that nearby piece?"
Real projects differ in every step. Some use light microscopes on expanded tissue instead of electron microscopes. Others use X-rays, which can image tissue without cutting it. Steps also overlap: synapses can be found before proofreading ends, and proofreading continues for years after the first wiring diagram is released.
Check yourself
-
Why does connectomics use an electron microscope rather than an ordinary light microscope?
Answer
Many neuron branches are thinner than about 200 nm, the smallest detail an ordinary light microscope can show. An electron microscope shows details of a few nanometers, so it can follow thin branches and see synapses.
-
Put these in order: find synapses, segment, proofread, stack into a 3D image.
Answer
Stack into a 3D image, segment, proofread, find synapses. (In practice, synapse finding and proofreading often overlap.)
-
A program segments a volume with 100,000 neuron pieces and makes an error on just 1% of them. How many errors must people find?
Answer
1% of 100,000 is 1,000 errors. That is why software that points people to likely errors saves so much time.
Go further
Optional background. This overview covers what you need here.
-
A talk for a general audience on why anyone would map every connection in a brain, and how it is done.
-
A short history of neuron-by-neuron maps, from the first whole worm to whole fly brains, The article's later "Public datasets" section links to the data.
-
A public connectome of a whole adult fruit fly brain, proofread by many people; you can browse its neurons online.
Where this shows up
No unit or explanation builds on this one yet.